Table 3.
The application of AI in upper-airway obstruction assessment.
| Author (Year) | Data Type | Dataset Size (Training/Test) |
Algorithm | Purpose | Performance |
|---|---|---|---|---|---|
| Shen et al. (2020) [109] | Lateral cephalograms | 488/116 (additional 64 images than validation set) |
CNN | Adenoid hypertrophy detection | Classification accuracy: 95.6%. Average AN ratio error: 0.026. Macro F1 score: 0.957. |
| Zhao et al. (2021) [110] | Lateral cephalograms | 581/160 | CNN | Adenoid hypertrophy detection | Accuracy: 0.919. Sensitivity: 0.906. Specificity: 0.938. ROC: 0.987. |
| Liu et al. (2021) [111] | Lateral cephalograms | 923/100 | VGG-Lite | Adenoid hypertrophy detection | Sensitivity: 0.898. Specificity: 0.882. Positive predictive value: 0.880. Negative predictive value: 0.900. F1 score: 0.889. |
| Sin et al. (2021) [112] | CBCT | 214/46 (additional 46 images than validation set) |
CNN | Pharyngeal airway segmentation | Dice ratio: 0.919. Weighted IoU: 0.993. |
| Leonardi et al. (2021) [113] | CBCT | 20/20 | CNN | Sinonasal cavity and pharyngeal airway segmentation | Mean matching percentage (tolerance 0.5 mm/1.0 mm): 85.35 ± 2.59/93.44 ± 2.54 |
| Shujaat et al. (2021) [114] | CBCT | 48/25 (additional 30 images than validation set) | 3D U-Net | Pharyngeal airway segmentation | Accuracy: 100%. Dice score:0.97 ± 0.02. IoU: 0.93 ± 0.03. |
| Jeong et al. (2023) [115] | Lateral cephalograms | 1099/120 | CNN | Upper-airway obstruction evaluation | Sensitivity: 0.86. Specificity: 0.89. Positive predictive value: 0.90. Negative predictive value: 0.85, F1 score: 0.88. |
| Dong et al. (2023) [116] | CBCT | A total of 87 | HMSAU-Net and 3D-ResNet | Upper-airway segmentation and adenoid hypertrophy detection | Segmentation: Dice value: 0.96. Diagnosis: accuracy: 0.912. Sensitivity: 0.976. Specificity: 0.867. Positive predictive value: 0.837. Negative predictive value: 0.981. F1 score: 0.901. |
| Jin et al. (2023) [117] | CBCT | A total of 50 | Transformer and U-Net | Nasal and pharyngeal airway segmentation | Precision: 85.88~94.25%. Recall: 93.74~98.44%. Dice similarity coefficient: 90.95~96.29%. IoU: 83.68~92.85%. |
ROC, receiver operating characteristic; CBCT, cone-beam computed tomography; CNN, convolutional Neural Network; AN, adenoid–nasopharynx; IoU, Intersection over Union; HMSAU-Net, hierarchical masks self-attention U-net; 3D, three-dimensional.